Skip to content

Article · AI in production

A Data Lake Won’t Get AI Into Your Workflows, and a Thousand Pilots Won’t Either

Big companies keep starting AI with a multi-year data lake or a pile of disconnected pilots. Both stall for the same reason. Here’s what gets AI into production.

On this page: Two myths
Title card for Brian Cosmano’s article, A Data Lake Won’t Get AI Into Your Workflows, and a Thousand Pilots Won’t Either, with the Athena wordmark and advicebyathena.com.

Somewhere in your firm there is a data strategy deck with a lake on page three. Somewhere else, a dozen people are running their own AI experiments. Neither effort has reached the place where an advisor prepares for a client review, and the people funding both are starting to ask why. The two efforts rest on opposite myths, and both fail for the same reason.

Two myths that sound like opposites

The first myth says the key first step is a data strategy that gets all of the company’s data into a lake, warehouse, or lakehouse so AI has context. It has serious backing. McKinsey, the consulting firm, wrote on Sept. 15, 2023: “If your data isn’t ready for generative AI, your business isn’t ready for generative AI.” Gartner, the research firm, predicted on Feb. 26, 2025 that through 2026 organizations “will abandon 60% of AI projects unsupported by AI-ready data.” Piyush Gupta, chief executive of Singapore’s DBS Bank, told McKinsey on Feb. 10, 2025: “You have to be thoughtful about how you create and architect your data lakes.” He said the cleanup and migration took “more than ten years and about a 100-person data factory.” He also described building demonstrable use cases along the way to show value.

The second myth says to let every employee build their own small AI prototypes and expect them to add up to an enterprise capability. The results so far say otherwise. MIT’s NANDA initiative, a research group at the Massachusetts Institute of Technology, studied this in its July 2025 report “The GenAI Divide: State of AI in Business 2025.” It found 95 percent of organizations getting no measurable P&L return from generative AI. For custom or vendor-sold enterprise tools, 60 percent of organizations evaluated them, 20 percent reached pilot, and 5 percent reached production. The authors did not blame the models. They pointed to brittle workflows, a lack of contextual learning, and poor fit with day-to-day operations.

Bar chart from MIT NANDA’s July 2025 report, The GenAI Divide: of organizations that looked at custom or vendor-sold enterprise AI tools, 60 percent evaluated them, 20 percent reached pilot, and 5 percent reached production.

Data quality matters, and experiments have value. What fails is the sequence and the expectation of scale.

Why the lake-first sequence does not hold

Gartner’s 2024 roadmap “AI-Ready Data Essentials” says AI-ready data “must be representative of the use case.” It adds that this readiness “is not something you can build once and for all, nor that you can build ahead of time for all your data.” The first stage is to assess data needs depending on the AI use cases.

The record on big, ahead-of-time data programs is worse than the pitch. Gartner warned on July 28, 2014, in “Beware of the Data Lake Fallacy,” that without descriptive metadata and a way to maintain it, “the data lake risks turning into a data swamp.” Qlik, a data analytics vendor, quoted a Sept. 26, 2017 Gartner note in a press release: “Through 2018, 90 percent of deployed data lakes will be rendered useless, as they’re overwhelmed with information assets captured for uncertain use cases.” That was a prediction, not a measured failure rate, but the diagnosis is specific: data captured for uncertain use cases.

McKinsey reported on Aug. 1, 2018 that one industrial company invested $15 million in big data infrastructure and did not start to see a payoff for at least three years. On Feb. 28, 2025, McKinsey wrote that most banks made significant data-transformation investments “in the past five to ten years,” and that “most banks either could not finalize them or could not realize their expected impact.” Unfinished programs left fragmented warehouses and lakes and “elevated costs from maintaining multiple complex environments.”

McKinsey’s March 4, 2024 guidance on generative AI names the alternative. Companies that do this well “tie their data quality and augmentation efforts to the specific AI/gen AI application and use case.” In the same piece it says “you don’t need this data foundation to extend to every corner of the enterprise.”

Why a thousand pilots do not scale

A pilot proves that something can work for one person on one afternoon. A workflow has to work for a team, every week, with the same inputs and the same checks.

The experiments are everywhere. Microsoft and LinkedIn’s Work Trend Index, a May 2024 survey of 31,000 knowledge workers in 31 countries, found that 75 percent use generative AI at work, and 78 percent of those users bring their own AI tools. Salesforce, the enterprise software company, found in a November 2023 survey of more than 14,000 workers that more than half of generative AI users at work had no formal employer approval, and nearly 7 in 10 had never received training. Activity that wide, with no shared workflow behind it, rarely turns into production systems.

IDC, the market research firm, in research commissioned by Lenovo, the computer maker, and reported by the trade publication CIO.com on March 25, 2025, found that for every 33 AI proofs of concept a company launched, only four reached production. IDC’s authors blamed low readiness in data, processes, and IT. Ashish Nadkarni, a group vice president at IDC, said many proofs of concept are “born at the board level” and are “highly underfunded or not funded at all.” Brian Jackson, a principal research director at Info-Tech Research Group, an IT research firm, offered the fair counterpoint: “even the failures are not failures if there are good lessons learned.”

Chart from IDC research commissioned by Lenovo, reported by CIO.com on March 25, 2025: a 33-dot grid in which only four dots are highlighted, showing that for every 33 AI proofs of concept, only four reached production.

McKinsey’s AI survey, fielded in July 2024 among 1,491 respondents and published March 12, 2025, found more than 80 percent not seeing a tangible impact on enterprise-level earnings before interest and taxes (EBIT) from generative AI. Only 21 percent of those reporting gen AI use had “fundamentally redesigned at least some workflows.” Of 25 attributes McKinsey tested, workflow redesign had the largest effect on EBIT impact. Its State of AI survey of 1,993 respondents, fielded from June 25 through July 29, 2025 and published Nov. 5, 2025, found 88 percent using AI in at least one function, yet nearly two-thirds had not begun scaling it across the enterprise. The same survey found use-case-level cost and revenue benefits, and 64 percent of respondents said AI was enabling innovation.

What gets AI into production

Three-column comparison of AI approaches: lake first, a thousand pilots, and use case first, with five steps for getting a use case into production and a note that MCP connects AI to the systems where data already lives.

Start with one use case that has an owner and a place in someone’s week. Pick the workflow it lives in, such as preparing for a client review or drafting the follow-up after one. List the data that job needs, and find where that data already lives. Fix the quality of that data for that job, and no more. Then build the AI step into the workflow, so the person doing the work meets it in the tool they already use.

The connection step used to be the hard part. Model Context Protocol, or MCP, is an open way for AI tools to connect to the systems where data already lives, so teams do not have to move everything into one lake first. Amazon Web Services, the cloud arm of Amazon, described the pattern in a Sept. 4, 2026 post on its Big Data Blog: instead of moving all data to one place, “let AI agents talk directly to the systems where data lives. Model Context Protocol (MCP) makes this possible.” A lake can still earn its place. Ernst & Young (EY), the professional services firm, wrote in its 2026 paper “Model context protocols vs. enterprise data platforms”: “The EDP governs the data; MCP governs the AI’s access to that data.”

Each use case that ships this way leaves behind cleaner data, a working connection, and a team that has seen the result. The next use case starts from there. Scale happens one workflow after another. It does not require a decade of preparation, and it does not come from a hundred scattered experiments that never meet.

That is the approach Athena takes. Athena is the AI-native platform built for how advisors work, and its MCP connection is live today, so a firm’s AI tools can work with Athena where the data already lives. We also help firms with their AI implementation, one workflow at a time.

Sources

Written by
Brian Cosmano, founder and CEO of Athena Advisory
Published
November 16, 2026
Topic
AI in production
Book a demo

See it on a book like yours.

Book a demo and we’ll walk a household from the client app to Mission Control.

Book a demo
Written by
Brian Cosmano, founder and CEO of Athena Advisory
Published
November 16, 2026
Topic
AI in production
Book a demo